Software development is changing. Tool calling, inference scaling and RL with Verifiable Rewards have combined over the past year to enable agent harnesses like Claude Code which can reliably navigate, modify and contribute to large codebases.
LLMs scale amazingly well with the amount of training data you throw at them. But I’ve been thinking about how to build tools that work alongside the characteristics of LLMs rather than language models needing to learn how to work with existing human-centric tools during training.
I have a hunch that a programming environment built around the strengths and limitations of autoregressive LLMs can lead to cheaper and higher-quality agent-powered development. How could we prove out that hypothesis? One would first need to design a language that aligns with how LLMs “think”. What would such a language look like? In this post I put forward some ideas for a language called Markov that I think would fit the bill.